Modeling Neural and Wearable Sensor Signals for Human-centric Applications

Abstract

This thesis explores novel approaches for modeling neural-physiological, peripheral-physiological, and wearable sensor signals to improve human-centric applications, addressing key challenges such as inter-participant variability, data scarcity, and context-specific feature representation. Human-centric applications prioritize user experience by leveraging physiological and behavioral signals to create adaptive and intuitive systems. The research is structured around three primary objectives: (i) Understanding neural-physiological signals in audio perception: This objective explores how neural signals, particularly EEG data, can provide insights into auditory perception. The research focuses on two key applications which are estimating cognitive load during sonification tasks and identifying songs using EEG signals. The goal is to uncover latent neural representations that capture discriminative features for these audio-related tasks. (ii) Addressing data scarcity in physiological signals: This objective tackles the challenge of limited datasets by generating high-quality synthetic data using advanced generative modeling techniques. These methods aim to augment small datasets while preserving the complexity and diversity of real-world physiological signals. (iii) Leveraging peripheral-physiological and wearable sensor signals: This objective investigates how data from wearable sensors (e.g., accelerometers, heart rate monitors) can provide deeper human-centric insights. The research focuses on extracting robust, context-aware features for applications like activity recognition by integrating insights from neural signal modeling and generative approaches. By addressing key challenges in physiological signal modeling, this thesis lays a foundation for creating more intuitive, context-aware, and user-centered applications. Ethical considerations were rigorously followed throughout the research process to ensure data integrity.